MCW Framework — Draft Paper Outline¶
Working title: The Meta-Context Window: A Coordination Framework for Human–AI Collaboration Failures
Intended venue: CHI, CSCW, or arXiv preprint (cs.HC / cs.AI)
Paper type: Framework / Conceptual Contribution
Target length: 6–10 pages (short paper) or extended abstract (4 pages)
Status: Outline only · Evidence: L0–L1. Evidence notes indicate what currently exists vs. what is still needed.
Epistemic Posture¶
This paper is a framework contribution, not an empirical validation. The recognized genre at CHI/CSCW introduces: - a new vocabulary for a previously unnamed phenomenon - a falsifiable taxonomy of failure modes - a designed experimental protocol
This posture allows rigorous claims without overclaiming results that do not yet exist. The abstract and introduction must make this framing explicit.
Abstract (Draft)¶
Failures in human–AI collaboration are routinely attributed to model capability — hallucination, misunderstanding, misalignment. We propose that a significant and underexplored class of these failures is better explained as coordination failures: breakdowns in the shared context that emerges between a human and an AI during interaction.
We introduce the Meta-Context Window (MCW) — a formal construct describing the dynamically maintained coordination state that arises when a human cognitive system (HCW) and an artificial cognitive system (ACW) exchange Information Units (IUs) through a constrained channel. MCW is neither stored in the AI nor in the human; it exists only through active, bidirectional exchange.
We present: (1) a formal definition of MCW and its component constructs; (2) a taxonomy of six failure modes with observable early-warning signals; (3) a set of repair operations derived from information-theoretic principles; and (4) five lightweight, falsifiable toy experiments designed to probe coordination dynamics without requiring instrumented systems or benchmarks.
The MCW framework does not compete with alignment, safety, or prompt engineering approaches. It offers a complementary lens: a way to diagnose coordination failures that occur even when models are capable, prompts are well-formed, and intentions are good.
1. Introduction¶
Key claim: Many human–AI collaboration failures are coordination failures, not capability failures.
Argument structure: 1. Increasing model capability and context window size have not uniformly improved collaboration quality 2. Current tooling (RAG, prompt engineering, fine-tuning) optimizes the AI's context window in isolation 3. The human has a context window too — and it is not modeled 4. Together, human and AI form a coupled system whose emergent coordination state has no name, no repair protocol, and no diagnostic framework 5. MCW names that state and provides the beginnings of a systematic treatment
Evidence available now: - Practitioner observation: expert AI users manage context instinctively, producing better outcomes — but this folk knowledge is tacit and non-transferable - "Lost in the middle" and maximum effective context window research shows that ACW size ≠ ACW effectiveness - Human–human communication research (repair theory, common ground) provides precedent for this class of failure
Evidence still needed: - Controlled studies demonstrating MCW degradation independently of model capability - Quantitative proxy metrics (repair cost, drift rate) correlated with outcome quality
2. Background and Related Work¶
Key claim: MCW is not captured by existing frameworks; it occupies a gap between HCI, CSCW, information theory, and AI alignment.
2.1 Human–Computer Interaction (HCI)¶
HCI focuses on usability, affordances, and task completion. MCW addresses coordination dynamics that persist even when interfaces are well-designed.
2.2 Computer-Supported Cooperative Work (CSCW)¶
CSCW studies shared artifacts and breakdowns in collaboration. The IU model and failure taxonomy are substrate-independent — applicable to human–human, human–AI, and AI–AI systems without presupposing social structure — while the MCW construct itself is canonically scoped to HCW–ACW coupling (Constitution v1.1, Article III), with generalizations as declared extensions.
2.3 Common Ground Theory (Clark & Brennan, 1991)¶
Common ground describes the mutual knowledge, beliefs, and assumptions shared by communication partners. MCW extends this to non-human actors and formalizes it through IU exchange and information-theoretic degradation.
Key distinction: Common ground is a property of shared beliefs; MCW is a property of shared coordination state — which includes salience, timing, and repair capacity, not just propositional content.
2.4 Information Theory¶
Shannon entropy provides the intuition for MCW degradation — not its mathematical grounding, since no random variable over the coordination state is defined (see the glossary's Entropy status). MCW does not attempt to formalize cognition in Shannon-theoretic terms; it borrows the intuition of information loss under constrained transmission and says so.
2.5 Alignment, Safety, and Prompt Engineering¶
These approaches optimize within the ACW. MCW addresses the between — the coupling layer that is currently unmodeled.
Evidence available now:
- docs/related_work.md in this repository
Evidence still needed: - Formal literature review (currently narrative) - Explicit comparison to common ground theory, distributed cognition (Hutchins), and situated action (Suchman)
3. The MCW Framework¶
Key claim: MCW is a formally definable, measurable construct — not a metaphor.
3.1 The Three Context Windows¶
Define HCW, ACW, and MCW. Establish that: - HCW is continuous, non-enumerable, advances between turns - ACW is discrete, bounded, updates only on input - MCW emerges from their coupling — it is not stored anywhere
3.2 Information Units (IUs)¶
Define IU as the minimal transferable element that can influence coordination state. Establish substrate-independence and representation-dependence.
Map IUs to existing AI concepts:
| AI Concept | IU Interpretation |
|---|---|
| Tokens / embeddings | IU representations |
| RAG | IU retrieval and reinjection |
| Prompt engineering | IU bundling |
| Context drift | IU entropy accumulation |
| Expert performance | Better IU stewardship |
3.3 MCW as a Coupling Function¶
Present the formal definition:
MCW = f(HCW₁…HCWₙ, ACW₁…ACWₘ, T, C)
The notation is informal [L0] — no codomain or functional form is defined, and the HCW is not enumerable; the paper must present it as a mnemonic for the qualitative properties (emergent, bidirectional, temporal, lossy, not a sum or product), not as mathematics (see the notation status in the glossary MCW entry). Explain why MCW is not a sum or product. Introduce the five-stage IU flow model (selection → encoding → transmission → decoding → integration).
Naturalistic micro-example (IU flow failure, Stage 3→4):
During construction of this framework's documentation site, icon shortcodes (e.g., :material-book-alphabet:) were authored correctly but rendered as raw text for all readers. The cause was a missing rendering extension — a hidden variable (constraint opacity) invisible to both the author and the reader. The repair operation was synchronization: making the constraint visible and resolving it. This is a single naturalistic observation with no experimental controls. Its value is illustrative, not evidential: it demonstrates that the IU flow model's failure modes are recognizable in routine technical work, not only in designed experiments. Epistemic weight: Layer 0 illustration only.
3.4 System Prompts as Initialization Artifacts¶
Establish that a system prompt is a static IU bundle — not an MCW participant. It biases early coordination but cannot maintain, monitor, or repair MCW.
Evidence available now:
- Full framework formalization in README.md and docs/glossary.md
Evidence still needed: - Diagrams illustrating HCW/ACW/MCW relationships and IU flow - Formal notation review by an information theorist
4. Failure Taxonomy¶
Key claim: MCW degrades through six identifiable, substrate-independent failure modes with observable early-warning signals.
| Failure Mode | Definition | Early Signal | Falsification |
|---|---|---|---|
| Drift | Silent context divergence | "That's not what I meant" arrives late | Interaction with explicit sync checkpoints shows equivalent drift |
| Asymmetric State Advancement | One actor advances off-turn without externalizing | Frustration at repetition; reset makes it worse | Off-turn reflection externalized immediately eliminates lag |
| False Alignment | Apparent agreement masks divergent interpretation | Confidence rises while accuracy drops | Verification checks expose misalignment before it compounds |
| Overcompression | Premature summarization destroys distinctions | Summary feels "off" but is unobjectionable | Delayed summarization preserves edge cases |
| Constraint Opacity | Hidden variables block repair | Repair attempts target wrong cause | Constraint disclosure reduces misattribution |
| Repair Suppression | Clarification signals are penalized or discouraged | Questions stop; errors recur | Explicit repair permission increases alignment over time |
Evidence available now:
- Failure taxonomy in README.md
- Toy experiment designs in experiments/toy_experiments.md
Evidence still needed: - Observed instances from real interaction logs (with consent) - Inter-rater reliability for failure mode classification
5. Repair Theory¶
Key claim: Repair is a first-class coordination primitive, not a social nicety — and can be formalized as a set of IU operations.
Present the five canonical repair operations: - Re-grounding - Decompression - Re-weighting - Disambiguation - Synchronization
Establish the falsifiable ordering claim [L0]: repair cost is non-decreasing in discovery lag (the rubrics' late-discovery measure). Earlier drafts asserted "exponentially cheaper" early repair — a functional form with no supporting data, withdrawn per Article IV. The immune-response and error-correction parallels are motivating analogies and must be presented as such.
Evidence available now:
- Repair operations in README.md and docs/glossary.md
- Qualitative observation from extended practitioner use
Evidence still needed: - Empirical repair cost curves - Comparison of repair cost pre- vs. post-MCW-aware interaction protocols
6. Toy Experiments¶
Key claim: MCW dynamics can be probed with lightweight, falsifiable experiments that require no instrumented systems or benchmarks.
Design Philosophy¶
- Coordination over capability: hold task difficulty constant; vary coordination conditions
- Qualitative-first: early signals are experiential and behavioral
- Comparative: results are meaningful relative to a baseline
- Repair-aware: treat repair latency and cost as first-class outcomes
Measurement Proxies (0–3 ordinal scales)¶
- H (MCW Health): perceived shared understanding
- R (Repair Cost): effort to realign
- D (Drift Rate): speed of divergence
- M (Misattribution): tendency to blame agent capability
Experiment Summary¶
| # | Name | Type | Failure Mode | Hypothesis |
|---|---|---|---|---|
| 1 | False Alignment Injection | Human ↔ Human | False Alignment | Ambiguous agreement language increases H↓, R↑, D↑ |
| 2 | Asymmetric State Advancement | Human ↔ AI | Asymmetric Advancement | Off-turn reasoning without externalization causes phase lag (R↑, M↑) |
| 3 | Overcompression Damage | Human ↔ AI | Overcompression | Premature summarization produces late, hard-to-debug failure |
| 4 | Constraint Opacity Stress Test | Human ↔ AI | Constraint Opacity | Opaque constraints increase H↓, M↑ vs. disclosed constraints |
| 5 | Repair Signal Suppression | Human ↔ Human | Repair Suppression | Suppressed clarification accelerates drift; late correction is costly |
| 6 | Drift Accumulation (declared extension) | Human ↔ AI | Drift | Without re-grounding checkpoints, goal representations diverge measurably and observably |
Full specifications, graded outcome interpretations, and cross-experiment analysis in experiments/toy_experiments.md.
Evidence available now: - Complete experiment designs with hypotheses and graded outcomes
Evidence still needed: - Pilot runs and recorded observations - Inter-rater reliability for proxy scores
7. Limitations¶
This section is critical for credibility with reviewers.
- No quantitative validation yet. Toy experiments are qualitative and observational. Claims are hypothesis-level, not evidence-level.
- Non-deterministic systems. Both human cognition and LLM behavior involve emergent properties that resist controlled isolation. Independent variable control is genuinely difficult.
- Measurement proxies are ordinal. H, R, D, M are coordination observables, not interval-scale metrics. Cross-study comparison is limited until instruments are validated.
- Coordination vs. capability confound. Separating MCW failures from genuine model capability failures requires experimental designs not yet implemented.
- Single-interaction focus. The framework currently addresses dyadic human–AI interaction. Multi-agent and organizational MCWs are theoretical extensions, not tested claims.
- Researcher positionality. The framework was developed through practitioner experience. Formal replication by independent researchers is needed.
8. Future Work¶
- Pilot studies using the toy experiment protocols with diverse participants
- Quantitative proxy instrument development and validation
- Hugging Face Space test bed for systematic A/B comparison (MCW-aware vs. baseline prompts)
- Extension to multi-agent and organizational MCW dynamics
- Integration with existing common ground and distributed cognition theory
- Application to non-AI domains (human–human, organizational, policy)
9. Conclusion¶
Key claim: MCW provides a minimal, falsifiable framework for studying coordination failures in human–AI collaboration — a class of failure that current tooling and research systematically misattributes to capability.
Restate: - The naming of MCW is itself the first repair operation: unnamed variables cannot be systematically reduced - The framework is designed to fail quietly if not useful — experiments should produce null results if MCW is not a real factor - The goal is not to replace existing approaches but to add a missing layer of analysis
Artifact Checklist (pre-submission)¶
| Artifact | Status |
|---|---|
| README with full framework | ✅ Complete |
| Canonical glossary | ✅ Complete |
| Toy experiments with graded hypotheses | ✅ Complete |
| Related work positioning | ✅ Complete |
| Test bed documentation | ✅ Complete |
| CITATION.cff with ORCID | ✅ Complete |
| Paper outline | ✅ This document |
| Framework diagrams (HCW/ACW/MCW, IU flow) | ✅ Complete (5 Mermaid diagrams in docs/diagrams.md) |
| Pilot study observations | ⬜ Needed |
| System prompt derivation doc | ✅ Complete |
| MCW Constitution | ✅ Complete (docs/constitution.md) |
| Hugging Face Space | ⬜ Planned |
| Formal literature review | ⬜ Needed |
This outline is a living document. Section order and emphasis should be revisited after pilot studies and peer feedback.